Machine Learning approaches for damage detection strategy in Structural Health Monitoring: application to experimental data

Structural Health Monitoring (SHM) is a discipline born during the last decades of the 20th century with the aim of implementing damage detection strategies in many engineering fields. These processes rely on structural observations that can be carried out through periodical dynamic response measurements coming from sensors installed on the structures. Datasets regarding a broad range of the lifetime of the structure are often required in order to implement efficient monitoring systems. Furthermore, analyses on the data should be fine evaluated in order to integrate and discern from the periodic or continuous measurements the noise (environmental) ones that may otherwise affect the results in a strong way. The first approach towards SHM is constituted by the model-driven methods: they can be developed starting from an updating of a Finite Element Model (FEM) by means of an inverse approach. The key concept relies on the definition of the modal parameters of the structure in order to minimize the discrepancy between the actual results and the prediction ones. The main problem related to the model-based approach concerns the noise data implementation, in which the environmental ones fall into. Instead, other SHM methods rely on the data-driven approach, which needs a big amount of data coming from permanent monitoring systems or from simulations aimed at describing accurately the actual behavior of the structure. Unlikely the first approach, data-driven methods can predict environmental variations by exploiting the algorithms capabilities to discern through data the real pathological structural states from the fictitious ones. This thesis is mainly focused on developing analyses embedded in the data-driven methods, more precisely focusing on exploiting a correlation between the amount of available data (or a limited part of them) and certain diagnostic characteristics of the structure. In the recent years, Machine Learning (ML) techniques have been used in order to enhance their properties related to extraction of information through a decision making approach: within a dataset, it is actually possible to “train” an algorithm and then test its capabilities to recognize a statistical pattern of information related to a possible damage in the structure. Transfer Learning techniques will be exploited to the behavior of different materials framed structures without having direct data from those, but only transferring knowledge between machine learning algorithms with damage detection objective.